An agent learns by doing. Somewhere.
A model can read every description of a task ever written and still not do it. What closes that gap is acting and finding out — and that has to happen somewhere. That somewhere is either a world you can simulate or a place where the work actually happens. benned licenses both from the people who own them, and operates them.
A place an agent acts in and gets consequences back, rather than a dataset it reads. It can be simulated — a licensed commercial game, or a world built for the purpose — or physical, meaning a real place of work where what skilled people do is recorded precisely enough to train on. benned licenses both from whoever owns them, makes them usable for a lab, and runs them. The owner keeps the relationship and earns; the lab gets something its competitors cannot simply buy a copy of.
| Simulated | Physical | |
|---|---|---|
| What it is | A game or built world the agent plays | A working site where people are recorded |
| How the agent uses it | Acts, fails, resets, retries | Learns from what a skilled human did |
| Can be reset and seeded | Yes | No |
| Faster than real time | Often | Never |
| Teaches best | Planning, sequence, consequence | Contact, grip, deformable material |
| Who owns the asset | A game studio | A business and the people who work there |
| What benned does | Makes it agent-ready and hosts it per lab | Specifies, records, clears and delivers it |
| Live today | Yes | No — an open programme |
The transaction is the same one twice.
Someone owns something a lab cannot get to. We make it usable and operate it. The lab pays, and the owner earns without handing over the customer. Whether the thing owned is a decade-old simulator or a bakery at five in the morning changes the engineering completely and the structure not at all.
- 1
Licensed game worlds
Commercial games made agent-ready for one lab: headless, with resets, seeds and scored tasks, on separate hardware. What a lab gets.
- 2
benned Worlds
Environments we build ourselves for long-horizon planning, starting with BuildSite. What we are building.
- 3
Recorded work
Skilled people doing their actual job, captured to a published standard, licensed by the business and consented to by the worker. The standard we hold.
- 4
Operation, in both cases
Nothing here is a brokerage. The environment is run, the capture is staffed, and the owner never has to become a technology company. What it takes from an owner.
Nobody knows yet which of these wins.
If world models turn out to learn actions from ordinary video, recorded demonstration gets cheap and simulated environments stay essential. If demonstration data turns out to be the thing that makes robots work, places of work become the scarce asset and simulators are a supporting act.
We are not in a position to know which, and neither is anyone selling you a confident answer. Holding both is not indecision — it is the only honest response to a question the field has not settled, and it costs us the focus that a single bet would buy.
What has actually happened so far- Licensed, never scrapedBoth kinds rest on a signed grant from whoever owns the thing. That is the whole company, in one clause.
- The owner keeps the customerA studio or a business contracts with the lab. We operate, and earn alongside rather than in between.
- One lab at a timeNo catalogue and no pooled data. The aggregation that would make this scale is the thing labs refuse.
One half of this is live. The other half is a standard.
The game side is the working business: licensed simulators, made agent-ready, operated per lab. The physical side has a written capture standard, a legal structure and no signed site. We have not recorded a working day for anyone.
That distinction matters more than it costs us to admit, because a company that blurs it is also a company that will blur what its data is worth. The standard is published in full so that the first business to sign one can see exactly what it is agreeing to, and so that a lab can hold us to it.
About environments
What is a training environment?
A place an agent acts in and gets consequences back, as opposed to a dataset it reads. It can be simulated — a commercial game or a purpose-built world — or physical, meaning a real place of work where the actions of skilled people are recorded. benned licenses both from whoever owns them and operates them for AI labs.
What is the difference between a simulated and a physical environment?
A simulated environment can be reset, seeded, run faster than real time and repeated exactly, which makes it the only practical way to train by trial and error at volume. A physical one cannot be reset or sped up, so it is recorded rather than played: what it gives you is the material behaviour and human dexterity that no simulator reproduces.
Why would one company do both?
Because the transaction is identical. In both cases someone owns something a lab cannot get to — a game, or a working bakery — benned makes it usable and operates it, the lab pays, and the owner earns without giving up the relationship. The two also hedge each other: if world models learn actions from video, environments still matter; if demonstrations win, places of work still matter.
Which should a lab buy first?
Whichever matches what its agents keep failing at. Failures of planning, sequence and consequence over long horizons point to a simulated environment. Failures of contact, grip and deformable material point to recorded human demonstration. Most serious agent programmes eventually need both.
Does benned operate physical environments today?
No. The game side is live and the physical side is a stated standard and an open programme, not a portfolio. benned has no signed capture site yet. Everything written here about physical capture describes what we hold ourselves to, and we would rather say that plainly than imply a track record we do not have.